An infrared target anti-interference tracking method based on feature segmentation tracking network

Abstract

The application provides an infrared target anti-interference tracking method based on a feature segmentation tracking network, relates to the image recognition technical field, and comprises three steps of constructing a neural network model, defining constraint conditions and judging a tracking state. Firstly, in the construction process of the feature segmentation tracking network model, the introduction of an attention mechanism and a multi-scale fusion module can accurately extract key information of a target and significantly suppress background interference. Secondly, a comprehensive loss function is adopted, a Soft-IoU loss, an energy loss and a hierarchical shape loss are combined, the segmentation result is constrained from three dimensions of overlap degree, energy and shape, the network prediction precision is significantly improved, and the tracking stability is maintained. Finally, in the target tracking process, the frame continuity and the shape feature matching are combined, the system can not only accurately judge whether the target is lost, but also quickly recover the tracking after the target is lost, and the target loss problem in the traditional method is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, in particular to an infrared target anti-interference tracking method based on a feature segmentation tracking network. BACKGROUND

[0002] Infrared target recognition and tracking technology has important application value in many fields such as military reconnaissance, security monitoring and other fields with high requirements for real-time and accuracy. However, the core challenge of this technology is how to ensure the continuous and stable tracking of the target in a complex interference scene. Complex background interference, target occlusion and motion blur and other factors often significantly affect the performance of the tracking algorithm. Existing researches mainly focus on improving the robustness of the tracking algorithm, especially by improving the feature extraction or motion prediction module to enhance the adaptability of the algorithm in the interference environment.

[0003] However, the current tracking method still has significant limitations in practical application. For example, when the target encounters partial occlusion or complex background interference, the existing algorithm often fails due to the insufficient recognition ability of the local features of the target, resulting in tracking drift or failure. Although some methods introduce segmentation networks to improve the positioning accuracy, such methods usually bring high computational complexity, which is difficult to meet the real-time tracking requirements.

[0004] Specifically, the interference sources in the complex ground scene, such as high-intensity areas, target-like backgrounds, smoke, and ground obstructions, can significantly reduce the recognizability of the target. Traditional global feature matching methods are prone to failure in this case, resulting in incorrect target positioning. For example, although deep learning-based algorithms can extract rich semantic features, their ability to perceive the overall shape features or local details of the target is weak, which greatly reduces their tracking performance when the target is partially occluded. In order to solve these problems, some methods attempt to combine segmentation and tracking modules, but these methods mostly use serial architectures, lack deep interaction between the segmentation and tracking modules, and are difficult to balance accuracy and efficiency.

[0005] In addition, most of the existing technologies do not fully utilize the synergistic effect of spatial local features and temporal context information of the target. In the actual tracking process, the apparent changes of the target and the uncertainty of the motion increase the difficulty of stable tracking. In order to improve the infrared target tracking performance in complex scenes, it is necessary to develop a new network architecture that deeply integrates local feature segmentation and temporal tracking mechanisms, and has good anti-interference ability and real-time performance, in order to significantly improve the robustness and accuracy of target tracking in complex environments. SUMMARY

[0006] The present application aims to provide an infrared target anti-interference tracking method based on a feature segmentation tracking network to overcome the deficiencies in the above technology.

[0007] To achieve the above object, the technical scheme adopted by the present application is as follows:

[0008] The present application provides an infrared target anti-interference tracking method based on a feature segmentation tracking network, comprising the following steps:

[0009] S1, constructing a neural network model: inputting the previous several frames of real images containing target regions and the current frame of real image, sequentially performing inter-frame attention enhancement, inter-frame difference enhancement, gradient enhancement and spatial position enhancement on the input image to obtain the feature enhancement map of the current frame, sequentially performing down-sampling, feature perception and feature fusion on the feature enhancement map, and outputting the prediction probability map of the current frame;

[0010] S2, defining the constraint condition: generating a comprehensive loss function according to the Soft-IoU loss function, the energy loss function and the hierarchical shape loss function, and the comprehensive loss function is used to constrain the prediction probability map from the three dimensions of overlap, energy and shape;

[0011] S3, judging the tracking state: extracting several candidate regions in the prediction probability map, calculating the comprehensive score of each candidate region through the Soft-IoU loss function and the hierarchical shape loss function, and judging the tracking state of the target by the neural network model according to the comprehensive score of each candidate region and the feature degradation degree index.

[0012] Further, in S1, the specific steps of sequentially performing inter-frame attention enhancement, inter-frame difference enhancement, gradient enhancement and spatial position enhancement on the input image to obtain the feature enhancement map of the current frame are as follows:

[0013] Generating a blur mask based on the binary mask of the previous frame of real image;

[0014] Differencing the real images of adjacent frames to generate several inter-frame difference channels;

[0015] Performing gradient filtering in different directions on the current frame of real image to construct gradient channels;

[0016] Normalizing the plane coordinates of each pixel in the current frame of real image to obtain the horizontal coordinate channel and the vertical coordinate channel;

[0017] Performing channel splicing and convolution operation on the real image, the blur mask, the inter-frame difference channel, the gradient channel, the horizontal coordinate channel and the vertical coordinate channel to obtain the feature enhancement map of the current frame.

[0018] Further, in S1, the specific steps of sequentially performing down-sampling, feature perception and feature fusion on the feature enhancement map to output the prediction probability map of the current frame are as follows:

[0019] The residual enhancement and down-sampling of the feature enhancement map generate a shallow feature map, a middle-level first feature map, a middle-level second feature map, a deep-level first feature map and a deep-level second feature map;

[0020] The middle-level first feature map and the deep-level first feature map are respectively subjected to multi-scale hollow convolution to generate a middle-level perception feature map and a deep-level perception feature map;

[0021] The middle-level second feature map and the deep-level perception feature map after up-sampling are subjected to cross-gate fusion to generate a first feature fusion map;

[0022] The first feature fusion map and the middle-level perception feature map are subjected to cross-gate fusion to generate a second feature fusion map;

[0023] The second feature fusion map after up-sampling and the deep-level second feature map after two-stage up-sampling are subjected to cross-gate fusion to generate a third feature fusion map;

[0024] The shallow feature map after residual enhancement and the third feature fusion map after two-stage residual enhancement are subjected to cross-gate fusion to generate a fourth feature fusion map;

[0025] The fourth feature fusion map is subjected to two-stage residual enhancement and convolution operation to obtain a prediction probability map.

[0026] Further, in S1, the residual enhancement includes residual convolution and gradient attention enhancement, the gradient attention enhancement includes channel attention enhancement and spatial attention enhancement, and the expression is:

[0027]

[0028]

[0029]

[0030]

[0031] wherein, represents an enhanced object, represents a channel number, convolution is performed on and output as channels, represents the first batch normalization, represents the ReLU function activation, represents the second batch normalization, represents the residual convolution output result, represents the average pooling, represents the maximum pooling, represents the first full connection, represents the second full connection, This indicates that the Sigmoid function is activated. This indicates the results of channel attention enhancement. Indicates channel splicing. This indicates the results of spatial attention enhancement. This represents the residual adjustment term. Indicates to Perform residual enhancement and output as One channel.

[0032] Furthermore, in S1, the expressions for generating shallow feature maps, intermediate first-level feature maps, intermediate second-level feature maps, deep first-level feature maps, and deep second-level feature maps by performing residual enhancement and downsampling on the feature enhancement map are as follows:

[0033]

[0034]

[0035]

[0036]

[0037]

[0038] in, Represents the feature enhancement map, This indicates that the residual has been enhanced. This represents two-dimensional max pooling. This represents a shallow feature map. Indicates the number of shallow channels. This represents the first-level feature map in the middle layer. This represents the second-level feature map of the middle layer. Indicates the number of middle-layer channels. This represents the deep first-level feature map. Represents a deep secondary feature map. This indicates the number of deep channels.

[0039] Furthermore, the expression for cross-gating fusion is:

[0040]

[0041]

[0042]

[0043]

[0044] in, This represents the high-resolution feature maps that need to be fused. This represents the low-resolution feature maps that need to be fused. Indicates the number of projection channels. Represents a 1×1 convolution. This indicates the result of high-resolution feature fusion. This indicates the result of low-resolution feature fusion. This indicates that the Sigmoid function is activated. Indicates the gating coefficient. This indicates the features after gating. Indicates channel splicing. This represents a 1×1 convolution with batch normalization and ReLU activation. This represents a feature fusion map.

[0045] Furthermore, in S2, the expression for the comprehensive loss function is:

[0046]

[0047] in, , , These represent the Soft-IoU loss weights, energy loss weights, and layered shape loss weights, respectively. Indicates Soft-IoU loss. Indicates energy loss. Indicates layered shape loss, This indicates the overall loss.

[0048] Furthermore, in S2, the expression for the energy loss function is:

[0049]

[0050]

[0051]

[0052] in, Represents pixel coordinates, Indicates the current frame's true image in grayscale value at that location The binary mask representing the real image of the current frame is in The value at ∑ represents the summation of pixel values ​​at all locations in the image. Represents normalized real energy. Indicates the current frame prediction probability map in The value at that location, This represents the normalized predicted energy. Indicates the smoothing term. This indicates energy loss.

[0053] Further, in S2, the expression of the hierarchical shape loss function is:

[0054]

[0055]

[0056]

[0057]

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067] wherein, represents the value of the binary mask of the real image of the current frame at or the value of the predicted probability of the current frame after segmentation at Figure Two represents the normalized gray value of any pixel in the target region of the real image of the current frame, represents the area of the target region, represents the horizontal coordinate of the center of the target region, represents the vertical coordinate of the center of the target region, represents the normalized maximum radius, represents the normalized horizontal coordinate of the center, represents the normalized vertical coordinate of the center, represents the normalized radius of the pixel , represents the dimension sequence number of the feature vector, represents the th feature weight of the pixel with the normalized radius , represents the th eigenvalue extracted using ,​​ Indicates use Extracted feature vector of the current frame, Indicates use The calculated first eigenvalues ​​of order 1 Indicates use The calculated first eigenvalues ​​of order 1 Indicates the first Relative difference of each feature Represents the numerically stable term. These represent the loss weights for the first, second, third, and fourth layers, respectively. These represent shape loss at one layer, two layers, three layers, and four layers, respectively. This indicates the loss of layered shape.

[0068] Furthermore, the specific steps of S3 are as follows:

[0069] Binarization and connected component operations are performed on the predicted probability map to extract several candidate regions, as shown in the expression:

[0070]

[0071] in, Indicates the binarization threshold. This represents the predicted probability map for the current frame. This indicates an 8-neighbor connectivity operation. Indicates the first Candidate regions, Indicates the area of ​​the candidate region. Indicates the area threshold. Represents the set of candidate regions;

[0072] The comprehensive score for each candidate region is calculated using the Soft-IoU loss function and the hierarchical shape loss function. The expression for the comprehensive score is as follows:

[0073]

[0074] in, These represent the weights of the Soft-IoU loss score and the weights of the hierarchical shape loss score, respectively. This represents the target region in the prediction probability map of the previous frame. express Soft-IoU loss, Represents the binary mask of the real image in the current frame. express Layered shape loss, This represents the overall score of the candidate region;

[0075] When the target is not occluded, the hierarchical shape loss of at least one candidate region is less than or equal to the feature degradation degree index, and the candidate region with the minimum comprehensive score value is selected as the predicted target region;

[0076] When the target is occluded, the hierarchical shape loss of all candidate regions is greater than the feature degradation degree index, and the neural network model is used to re-track the target until the hierarchical shape loss of at least one candidate region is less than or equal to the degradation degree index, and the candidate region with the minimum hierarchical shape loss value is selected as the predicted target region.

[0077] The beneficial effects of the present application include:

[0078] The present application provides an infrared target anti-interference tracking method based on a feature segmentation tracking network, which includes three steps of constructing a neural network model, defining a constraint condition and judging a tracking state. First, in the construction process of the feature segmentation tracking network model, by introducing an attention mechanism and a multi-scale fusion module, the key information of the target can be accurately extracted, and the background interference can be significantly suppressed. The operations of inter-frame attention enhancement, inter-frame difference enhancement, gradient enhancement and spatial position enhancement enable the network to identify and extract the detailed features of the target in a complex interference environment, rather than relying only on global features. This process greatly improves the distinguishability of the infrared target, especially in the case of complex background and blurred target, effectively avoiding the confusion between the target and the background, and improving the tracking accuracy.

[0079] Secondly, a comprehensive loss function is used, which combines Soft-IoU loss, energy loss and hierarchical shape loss, to constrain the segmentation result from three dimensions of overlap degree, energy and shape, thereby significantly improving the shape retention ability of the target, especially in the case of complex situations such as target occlusion or background mutation, the stable shape of the target can be maintained and the shape distortion can be effectively suppressed. Each item in the comprehensive loss function accurately optimizes the positioning accuracy and shape consistency of the target from different dimensions, ensuring that the target is not affected by external interference during the segmentation tracking process.

[0080] Finally, in the target tracking process, the comprehensive score mechanism and the feature degradation degree index are used to accurately judge the tracking state of the target. By scoring and matching multiple candidate regions, the system can quickly recover the target tracking based on the historical information of the target in the case of partial occlusion or loss of the target. Combined with the inter-frame continuity and shape feature matching, the system can not only accurately judge whether the target is lost, but also quickly recover the tracking after the target is lost, avoiding the target loss problem in the traditional method.

[0081] Overall, the method provided by the application adopts a lightweight network design, significantly optimizes the computing efficiency, and ensures the real-time performance of the method in practical applications. At the same time, through the hierarchical feature extraction and dynamic fusion mechanism, the loss function design and the tracking state judgment mechanism, the false alarm rate and the risk of target loss are reduced, ensuring the stability and efficiency of the network tracking, and showing excellent robustness and practicality in complex infrared scenes in the fields of military reconnaissance, security monitoring and the like, providing a new solution for infrared target tracking technology, and having wide practical application value and important market prospect. BRIEF DESCRIPTION OF DRAWINGS

[0082] Figure One A flowchart of an infrared target anti-interference tracking method based on a feature segmentation tracking network provided by the application;

[0083] Figure Two A framework diagram of an infrared target anti-interference tracking method based on a feature segmentation tracking network provided by the application;

[0084] Figure Three A detailed architecture diagram of a feature segmentation tracking network provided by the application;

[0085] Figure Four A feature segmentation tracking network training effect change diagram;

[0086] Figure Five An anti-suspected target interference tracking effect diagram;

[0087] Figure Six A target occlusion and successful re-identification recovery tracking effect diagram. DETAILED DESCRIPTION

[0088] The application is based on the following software and hardware environment and parameter configuration to fully verify the performance of the provided method in complex interference scenarios:

[0089] 1) Parameter configuration

[0090] The basic network parameters are as follows: the local patch size (Local_size) is set to 32 to ensure effective local feature extraction of the infrared small target; the dilate kernel (dilate_kernel) is 3, which moderately expands the feature perception range while preserving the target core area; the Gaussian blur radius (blur_radius) is 3.0, which enhances the fault tolerance to small inter-frame displacement.

[0091] The training parameters are as follows: the batch size (batch_size) is 1 to adapt to small sample infrared data training; the number of training rounds (num_epochs) is 20 rounds, the learning rate (lr) is 2e-3, which ensures that the model is fully converged and the training is stable.

[0092] Tracking parameters: The target area threshold (area_thresh) is set to 5 to filter out small interference areas at the pixel level and accurately extract effective target candidate areas.

[0093] 2) Dataset Description

[0094] The dataset used in the experiment is generated by modifying the target templates of the VOT-TIR infrared tracking dataset. Artificially set interference elements such as occlusions and suspected targets to simulate strong interference environments in complex battlefield and security scenarios, to more comprehensively test the performance of the proposed algorithm in infrared target anti-interference tracking tasks.

[0095] 3) Hardware Environment

[0096] The computer configuration relied upon in the experiment is as follows: equipped with an 11th Gen Intel (R) Core (TM) i7-11800H @ 2.30GHz processor, 16GB of memory (15.8GB available); the graphics card is NVIDIA GeForce RTX 3060Laptop GPU (6GB video memory) and Intel (R) UHD Graphics (128MB video memory); the storage consists of 477GB SSD SAMSUNG MZVL2512HCJQ-00BL2 and 954GB SSD Colorful CN700 1TB PLUS, providing sufficient and high-speed storage support for data reading and writing and model training, ensuring efficient experiment development.

[0097] The present application provides an infrared target anti-interference tracking method based on a feature segmentation tracking network, aiming to solve the problems of target loss and tracking drift caused by target occlusion and background interference in infrared images. As shown in Figure One and Figure Two The method comprises the following steps:

[0098] S1, constructing a neural network model: input the first several frames of real images containing target regions and the current frame of real image, and sequentially perform inter-frame attention enhancement, inter-frame difference enhancement, gradient enhancement and spatial position enhancement on the input images to obtain the feature enhancement map of the current frame, and sequentially perform down-sampling, feature perception and feature fusion on the feature enhancement map to output the prediction probability map of the current frame.

[0099] Specifically, the constructed neural network model is a Feature Aggregation Segmentation Tracking Network (FASTNet) model, which includes an enhancement module, an encoding down-sampling module, a multi-scale perception module, an inter-layer fusion module, and an output module. After inputting several frames of real images containing target regions, the enhancement module performs a series of enhancement operations such as inter-frame attention enhancement, inter-frame difference enhancement, gradient enhancement, and spatial position enhancement to enhance the features of the input images, thereby solving the problems of weak target features and spatial position ambiguity in infrared images and significantly improving the target recognition ability and tracking accuracy. After feature enhancement, the encoding down-sampling module extracts low-level detail features and high-level semantic information of the enhanced feature images, realizing the integration of feature extraction, attention enhancement, and resolution reduction. Then, the multi-scale perception module extracts features of different scales to capture small-scale details and large-scale context, enhancing the network's ability to recognize targets of different sizes. Next, the inter-layer fusion module dynamically fuses features of different resolutions, enabling the network to more comprehensively understand the spatial structure of the target and maintain the continuity of information flow, ensuring stable tracking of the target at different scales and resolutions. Finally, the output module enhances the feature fusion image and outputs the prediction probability map of the current frame. The prediction probability map can accurately represent the position and shape of the target and provide high-quality input for subsequent segmentation and tracking.

[0100] In summary, the feature segmentation and tracking integrated model constructed in step one adapts to complex infrared scenes, realizes accurate segmentation and continuous tracking of targets, and adopts a lightweight design to ensure efficient computing performance, adapt to real-time tracking requirements, and can be widely applied in military reconnaissance, security monitoring, and other fields, and exhibits excellent robustness and real-time performance in dynamic environments.

[0101] As shown in Figure Three , the specific steps of S1 are as follows:

[0102] S11: After sequentially performing inter-frame attention enhancement, inter-frame difference enhancement, gradient enhancement, and spatial position enhancement on the input image, a feature enhanced image of the current frame is obtained. The specific steps are as follows:

[0103] S111, inter-frame attention enhancement: a blur mask is generated based on the binary mask of the previous frame real image as an inter-frame attention clue to construct a time sequence association. The blur mask generation process is as follows:

[0104]

[0105] wherein, represents the binary mask of the previous frame real image, represents max-pooling dilation, in this application, the size of the pooling kernel is 3, represents Gaussian blur, in this application, the Gaussian blur radius is 3.0, represents the blur mask of the previous frame real image.

[0106] S112, inter-frame difference enhancement: difference between adjacent real images, generate several inter-frame difference channels as inter-frame gradient clues to further enhance the temporal correlation. The inter-frame difference channel generation process is:

[0107]

[0108] wherein, and represent the input of the first frame (previous several frames) and the first frame (current frame) real image, in this application, the real image size is 32x32, represents the frame order of the input real image, represents the total number of frames of the input real image, represents the generated inter-frame difference channel.

[0109] S113, gradient enhancement: gradient filtering of different directions on the current frame real image, construct gradient channels as current frame gradient clues, construct spatial gradient attention perception. The gradient channel generation process is:

[0110]

[0111] wherein, represents the current frame real image, represents the gradient operator direction, represents the gradient filtering of direction, represents the gradient channel of direction.

[0112] S114, spatial position enhancement: respectively normalize the plane coordinates of each pixel in the current frame real image to the [-1, 1] interval to obtain the horizontal coordinate channel and the vertical coordinate channel:

[0113]

[0114] wherein, represents the pixel coordinate, H and W represent the height and width of the input real image, in this application, , represents the horizontal coordinate channel, represents the vertical coordinate channel.

[0115] S115, coordinate convolution: channel splicing is performed on the real image, the blur mask, the several inter-frame difference channels, the gradient channels, the horizontal coordinate channels and the vertical coordinate channels, so that the network explicitly learns the target space position features, and then the channel number is changed through convolution operation, in the application, the channel number is changed to 16 through the convolution operation, and after batch normalization and activation, a feature enhancement map is obtained:

[0116]

[0117]

[0118] wherein, represents the channel number, in the application, , represents channel splicing, represents convolution and output as channels, represents the result after coordinate convolution, represents batch normalization, represents ReLU function activation, avoiding feature distribution deviation, represents the feature enhancement map.

[0119] Example: set the total number of input real images as 3, and take 0002, 0003 and 0004 frames in the infrared sequence image as the input real images, first calculate the target center coordinates of each frame of real image: 0002 frame is (147, 39), 0003 frame is (146, 38), and 0004 frame is (145, 37), after generating the horizontal coordinate channel and the vertical coordinate channel of the 0004 frame real image, the 0002, 0003 and 0004 frame real images, the blur mask (based on 0003 frame binary mask inflation + blur generation), the inter-frame difference channel (0003-0002 frame, 0004-0003 frame difference), and the four-direction gradient channel of the 0004 frame real image are spliced, and then the 0004 frame feature enhancement map with 16 output channels and 32*32 size is obtained through 3*3 convolution.

[0120] S12: residual enhancement and double-layer down-sampling are performed on the feature enhancement map to generate a shallow feature map, a middle-level first feature map, a middle-level second feature map, a deep-level first feature map and a deep-level second feature map, realizing the integration of feature extraction, attention strengthening and resolution reduction. The specific expression is:

[0121]

[0122]

[0123]

[0124]

[0125]

[0126] wherein, denotes a feature enhancement map, denotes residual enhancement, denotes two-dimensional max pooling, denotes a shallow feature map, denotes a shallow channel number, denotes a middle first-level feature map, denotes a middle second-level feature map, denotes a middle channel number, denotes a deep first-level feature map, denotes a deep second-level feature map, denotes a deep channel number.

[0127] Residual enhancement includes residual convolution and gradient attention enhancement, and gradient attention enhancement includes channel attention enhancement and spatial attention enhancement, and the expression is:

[0128]

[0129]

[0130]

[0131]

[0132] wherein, denotes an enhancement object, such as a feature enhancement map, a middle first-level feature map, or other feature maps that need to be enhanced by residual enhancement, denotes a channel number, denotes a convolution on and outputs channels, denotes the first batch normalization, denotes ReLU function activation, denotes the second batch normalization, denotes a residual convolution output result, denotes average pooling, denotes max pooling, denotes the first full connection, denotes the second full connection, denotes Sigmoid function activation, denotes a channel attention enhancement result, denotes channel splicing, denotes the spatial attention enhancement result, denotes the residual adjustment term, which is adapted by 1x1 convolution when the input and output channels or resolutions are inconsistent, denotes the residual enhancement on and outputs as channels.

[0133] Example: input the channel number of S11 generated feature enhancement map of 16 and the size of 32x32 into the encoding down-sampling module: first, a residual enhancement is performed to obtain a shallow feature map with a size of 32x32; the shallow feature map is down-sampled to 16x16 by two-dimensional maximum pooling, and then two residual enhancements are performed to obtain a middle two-level feature map with a channel number of 32 and a size of 16x16; the middle two-level feature map is down-sampled to 8x8 by two-dimensional maximum pooling, and then two residual enhancements are performed to obtain a deep two-level feature map with a channel number of 48 and a size of 8x8, completing the encoding down-sampling.

[0134] S13: the multi-scale perception module includes a middle multi-scale dilated convolution (DilConv) ) perception module and a deep multi-scale dilated convolution (DilConv) ) perception module, the middle multi-scale dilated convolution (DilConv) perception module and the deep multi-scale dilated convolution (DilConv) perception module are respectively used for multi-scale dilated convolution on the middle one-level feature map and the deep one-level feature map to generate a middle perception feature map and a deep perception feature map to respectively capture small-scale details and large-scale context. The specific expression is:

[0135]

[0136]

[0137] wherein, denotes the dilated rate, small-scale target details are captured by large-scale background context is captured by denotes a convolution with a dilated rate denotes channel splicing, denotes convolution plus batch normalization and ReLU activation, denotes the middle perception feature map, denotes the deep perception feature map.

[0138] Example: the middle one-level feature map (32 channels, 16x16) generated by S12 is applied to ​​The perception module: 32-channel feature maps are generated by 3x3 convolution with the hole rate of 1, 2 and 4, respectively, and are spliced and projected into middle-level perception feature maps (32 channels, 16x16) by 1x1 convolution. The perception module: 32-channel feature maps are generated by 3x3 convolution with the hole rate of 1 and 2, and are spliced and projected into deep-level perception feature maps (32 channels, 8x8).

[0139] S14: Dynamic fusion of feature maps with different resolutions is realized based on the cross-gated fusion module. The expression of cross-gated fusion is:

[0140]

[0141]

[0142]

[0143]

[0144] wherein, represents a high-resolution feature map that needs to be fused, represents a low-resolution feature map that needs to be fused, represents the number of projection channels, represents 1x1 convolution, represents a high-resolution feature fusion result, represents a low-resolution feature fusion result, represents a Sigmoid function activation, used to compress the result to the interval [0, 1], represents a gating coefficient, used to reflect the correlation of feature maps with different resolutions. The closer the value of the gating coefficient to 1, the stronger the correlation of feature maps with different resolutions at the position, and the closer the value to 0, the weaker the correlation of feature maps with different resolutions, represents a feature after gating screening, represents channel splicing, represents 1x1 convolution plus batch normalization and ReLU activation, represents a feature fusion map.

[0145] In this application, four fusions need to be realized by the cross-gated fusion module, and the specific steps are as follows:

[0146] S141: Cross-gated fusion of middle-level secondary feature maps and up-sampled deep-level perception feature maps at H / 2 resolution is performed to generate a first feature fusion map, and the expression is:

[0147]

[0148] wherein, denotes up-sampling, denotes cross-gated fusion, denotes the first feature fusion map.

[0149] S142: cross-gated fusion of the first feature fusion map and the mid-level perception feature map at H / 2 resolution to generate a second feature fusion map, expressed as:

[0150]

[0151] wherein, denotes the second feature fusion map.

[0152] S143: cross-gated fusion of the up-sampled second feature fusion map and the two-level up-sampled deep second-level feature map at full resolution to generate a third feature fusion map, expressed as:

[0153]

[0154] wherein, denotes the third feature fusion map.

[0155] S144: cross-gated fusion of the residual-enhanced shallow feature map and the two-level residual-enhanced third feature fusion map at full resolution to generate a fourth feature fusion map, expressed as:

[0156]

[0157] wherein, denotes two-level residual enhancement, and the output channel is C 1, denotes the fourth feature fusion map.

[0158] Instance description: 1) H / 2 resolution fusion: the deep perception feature map generated by S13 is up-sampled, and the middle two-level feature map (32 channels, 16x16) generated by S12 is input into the cross-gate fusion module with the up-sampled deep perception feature map (32 channels, 16x16) to generate a gating coefficient (16x16), and the first feature fusion map (32 channels, 16x16) is obtained after weighting; 2) full resolution fusion: the deep two-level feature map generated by S12 is two-level up-sampled, and the second feature fusion map (32 channels, 32x32) after two-level up-sampling is fused with the deep two-level feature map (32 channels, 32x32) after two-level up-sampling to obtain the third feature fusion map (32 channels, 32x32); the residual enhancement is performed on the shallow feature map generated by S12, and the third feature fusion map (16 channels, 32x32) after two-level residual enhancement is fused with the shallow feature map (16 channels, 32x32) after residual enhancement to obtain the fourth feature fusion map (16 channels, 32x32).

[0159] S15: two-level residual enhancement and convolution operation are performed on the fourth feature fusion map to obtain a prediction probability map, and the expression is:

[0160]

[0161]

[0162] wherein, represents the result of two-level residual enhancement on the fourth feature fusion map, represents 1x1 convolution, represents Sigmoid function activation, represents the prediction probability map of the current frame, the size is consistent with the input real image, , represents the pixel confidence of the target area.

[0163] Instance description: the fourth feature fusion map (16 channels, 32x32) generated by S14 is two-level residual enhanced, and then 1x1 convolution is performed to output the prediction probability map (32x32) of the current frame. Taking frame 0005 as an example, the prediction probability Figure Two value of frame 0005 is segmented into a target area of 30, and the global center coordinates are (143.5, 37.66666666666667), which is consistent with the real image segmentation tracking result of frame 0005.

[0164] S2, define the constraint condition: generate a comprehensive loss function according to the Soft-IoU loss function, the energy loss function and the hierarchical shape loss function, and the comprehensive loss function is used to constrain the prediction probability map from the three dimensions of overlap, energy and shape.

[0165] Specifically, based on the feature segmentation and tracking network (FASTNet) constructed in S1, S2 sets training constraints to constrain the entire learning process in three dimensions of "overlap-energy-shape", forming a collaborative convergence drive for the prediction probability map. The constraint system takes a comprehensive loss function as the core: reasonably weighting the Soft-IoU loss, energy loss and hierarchical shape loss in the same objective function, so that the network synchronously optimizes the three key indicators of regional overlap, area consistency and shape consistency in each forward-backward iteration. The logical relationship of the three parallel constraints is clear: Soft-IoU faces the optimization of the intersection over union at the pixel set level, solving the root cause of "inaccurate segmentation"; the energy term faces the area registration of the overall scale of the target, resolving the scale mismatch of "area drift"; the hierarchical shape term faces the geometric structure of the shape fidelity, relieving the geometric instability of "shape distortion"; the three dimensions complement each other and balance each other, avoiding the model to achieve "partial type" convergence relying on a single indicator, and jointly constraining the robust and continuous tracking of the network in complex dynamic environments.

[0166] Further, in S2, the expression of the comprehensive loss function is:

[0167]

[0168] wherein, , , Soft-IoU loss weight, energy loss weight, and hierarchical shape loss weight are respectively represented by =0.4, , , Soft-IoU loss, which optimizes the overall overlap, energy loss, which constrains the target energy, hierarchical shape loss, which constrains the target shape structure, comprehensive loss.

[0169] Further, in S2, the Soft-IoU loss function is a commonly used segmentation loss function, which can solve the discrete and non-derivable problem of traditional IoU loss and directly optimize the intersection over union indicator. The expression of the Soft-IoU loss function is:

[0170]

[0171] wherein, pixel coordinates are represented by the value of the binary mask of the current frame real image at , target pixel, background pixel, Indicates the current frame prediction probability map in The value at ∑ represents the summation of pixel values ​​at all locations in the image. To represent a smoothing term and avoid a denominator of zero, in this application, , Indicates Soft-IoU loss. The value range is [0,1]. The smaller the value, the higher the overlap between the predicted mask and the actual mask.

[0172] Example illustration: such as Figure Four As shown, the Soft-IoU loss of the FASTNet model constructed by S1 gradually decreased with each iteration during the 20 training rounds: it was 0.9577 in Epoch 1, decreased to 0.5517 in Epoch 10, and converged to 0.0174 in Epoch 20. This indicates that the overlap between the predicted mask and the real mask gradually increased from low to 98.26% (1-0.0174), and the segmentation accuracy was significantly optimized.

[0173] Furthermore, in S2, the expression for the energy loss function is:

[0174]

[0175]

[0176]

[0177] in, Indicates the current frame's true image in grayscale value at that location Represents normalized real energy. This represents the normalized predicted energy. Indicates energy loss. The value range is [0,1]. The smaller the value, the closer the predicted area is to the actual area, which can effectively improve the accuracy of area prediction in occluded scenarios.

[0178] Example illustration: The energy loss of the FASTNet model constructed by S1 gradually converged during training: 0.8670 in Epoch 1 (the predicted area differed greatly from the actual area), decreased to 0.0566 in Epoch 8, and stabilized at 0 in Epoch 10 and thereafter (the predicted area perfectly matched the actual area); during the tracking phase, the target area in frames 0005-0022 was stable between 17 and 30 (e.g., 30 in frame 0005, 28 in frame 0010, and 17 in frame 0022), and the energy loss was less than 0.1, indicating that the area prediction accuracy met the requirements.

[0179] Further, in S2, the hierarchical shape loss function first extracts the feature vector of the infrared target in the discrete digital image through the steps of gray scale normalization, center distance normalization and feature vector extraction, and different dimensions of the feature vector represent different center hierarchical quantization features of the target; then, the relative difference measurement method is used to calculate the feature difference; finally, the hierarchical weight structure is designed according to the different center levels of the target, and the hierarchical shape loss is calculated. The hierarchical shape loss combined with the pyramid weight distribution mechanism can maintain the overall structure of the shape while considering the local detail features, and effectively improve the shape keeping ability of the target in complex scenes. The expression is:

[0180]

[0181]

[0182]

[0183]

[0184]

[0185]

[0186]

[0187]

[0188]

[0189]

[0190]

[0191]

[0192]

[0193] wherein, represents the value of the binary mask of the real image of the current frame at or the value of the predicted probability Figure Two of the current frame after segmentation at , represents the normalized gray value of any pixel in the target region of the real image of the current frame, represents the area of the target region, represents the horizontal coordinate of the center of the target region, represents the vertical coordinate of the center of the target region, represents the normalized maximum radius, represents the normalized horizontal coordinate of the center, and denotes the normalized vertical coordinate of the center, denotes the pixel denotes the normalized radius of the pixel, denotes the feature vector dimension sequence number, denotes the first feature weight of the pixel with normalized radius , denotes the first order feature value extracted using , denotes the current frame feature vector extracted using , denotes the first order feature value calculated using , denotes the first order feature value calculated using , denotes the first relative difference degree of the feature, denotes the numerical stability term, in this application, , denotes the first layer loss weight, the second layer loss weight, the third layer loss weight, and the fourth layer loss weight, respectively, denotes the first layer shape loss, describing the overall uniform feature difference of the target, the second layer shape loss, describing the internal feature difference under the quadratic function distribution of the target, the third layer shape loss, describing the peripheral feature difference under the cubic function distribution of the target, and the fourth layer shape loss, describing the edge feature difference under the quartic function distribution of the target, denotes the hierarchical shape loss. The smaller the value, the higher the consistency between the predicted shape and the real shape.

[0194] Example: Take frame 0076 as an example. The feature vector extracted from the binary mask of the real image of frame 0076 is:

[0195]

[0196] The feature vector extracted after the value segmentation of the predicted probability of frame 0076 is: Figure Two

[0197]

[0198] Calculate the relative difference degree: , , , , , , , ; further calculate the shape loss of each layer: ,​ 、 、 ; total hierarchical shape loss , which indicates that the predicted shape is highly consistent with the true shape. The hierarchical shape loss decreases from 1.0802 in Epoch 1 to 0.0288 in Epoch 10, and finally converges to 0, verifying the effectiveness of the shape constraint.

[0199] S3, judging the tracking state: extracting a plurality of candidate regions in the predicted probability map, calculating the comprehensive score of each candidate region through the Soft-IoU loss function and the hierarchical shape loss function, and judging the tracking state of the target by the neural network model according to the comprehensive score of each candidate region and the feature degradation degree index.

[0200] Specifically, on the basis of the predicted probability map output in S1, through the closed loop of “candidate generation-reliability measurement-state judgment-history update / re-identification”, the optimal target of the current frame can be stably given under the conditions of multiple candidates, occlusion and strong interference, and the temporal prior is updated accordingly. The core idea is: to establish the posterior comprehensive score with Soft-IoU and hierarchical shape loss in S2 as the main shaft, and to make state judgment combined with area constraint, so that spatial appearance consistency, geometric shape consistency and temporal continuity can work together in the same evaluation framework, and higher target distinguishability and trajectory consistency can be obtained on the premise of maintaining real-time, and systematic, interpretable and verifiable performance improvement can be formed for the four key problems of “inaccurate segmentation, area drift, shape distortion and occlusion loss”.

[0201] The specific steps of S3 are as follows:

[0202] S31: as shown in Figure Five , the predicted probability map is binarized and connected domain operation is performed to extract a plurality of candidate regions. When the area of the candidate region is less than the area threshold, it is judged as low-confidence interference and directly excluded. When the area of the candidate region is greater than the area threshold, the candidate region is retained to enter the subsequent scoring and measuring link, which can not only reduce the scoring calculation overhead, but also avoid introducing obvious false targets into the decision. The expression is:

[0203]

[0204] wherein, denotes the binarization threshold, denotes the predicted probability map of the current frame, denotes the 8-neighborhood connected operation, denotes the th candidate region, denotes the area of the candidate region, denotes the area threshold, denotes the candidate region set.

[0205] Example: 1) Normal tracking frame (such as 0005-0022 frame): prediction probability of 0005-0022 frame Figure Two After binarization, only one candidate region with area 17-30 (all greater than the area threshold) is extracted as the prediction target region; 2) Occlusion frame (such as 0023 frame): prediction probability of 0023 frame Figure Two After binarization, the candidate region area is 3 (less than the area threshold), and after filtering, there is no effective candidate region, triggering the occlusion judgment of S33; 3) Multiple interference frame (such as 0076 frame): prediction probability of 0076 frame Figure Two After binarization, 3 candidate regions can be extracted through connected component analysis, with areas of 24, 8, and 10 respectively. After filtering the interference with an area less than 5, 3 candidate regions are retained (area 24, 8, and 10 respectively) ), enter the posterior scoring stage.

[0206] S32: Calculate the comprehensive score of each candidate region through the Soft-IoU loss function and the hierarchical shape loss function, and evaluate the matching degree of each candidate region with the historical target. The expression of the comprehensive score is:

[0207]

[0208] wherein, Soft-IoU loss score weight and hierarchical shape loss score weight respectively, denotes the target region in the prediction probability map of the previous frame tracking success, denotes the Soft-IoU loss of , used to measure the inter-frame continuity, the calculation process is shown in S2, denotes the binary mask of the real image of the current frame, denotes the hierarchical shape loss of , used to evaluate the shape consistency, the calculation process is shown in S2, denotes the comprehensive score of the candidate region, the value range is [0, 2], The smaller the value, the higher the matching degree of the candidate region with the historical target.

[0209] Example: Taking 0076 frame as an example, the posterior comprehensive score calculation of the 3 candidate regions is as follows:

[0210] Candidate region (area 24, target center coordinates (138.625, 96.043)): , , comprehensive score ;

[0211] Candidate region (Area 8, target center coordinates (139.210, 95.112)): , Overall score ;

[0212] Candidate region (Area 10, target center coordinates (137.982, 96.875)): , Overall score .

[0213] Choose the one with the lowest overall score. As the predicted target region in frame 0076, i.e. "Selected candidate 1 with score 0.8504", it ensures the continuity and accuracy of the tracked target.

[0214] S33: Based on the comprehensive score and feature degradation index, determine the current tracking status and select the optimal candidate region. For example... Figure Five As shown, under normal tracking conditions, when there is At that time, among them, Select the feature degradation threshold. The candidate region with the smallest value is used as the tracking result for the current frame. and update historical target information. .

[0215] like Figure Six As shown, when all candidate regions The system determines that the target has been occluded or lost. At this point, a re-identification mechanism is triggered, recording the coordinates of the target center from the last successfully tracked target. And perform target verification and precise positioning: after triggering the re-identification mechanism, with Centered on the current frame, a preset search radius is established within the global image. A sliding window scan is performed within the range, with each scan window having the same size as the real image input for network training. For each scan window... The data is input into the FASTNet network for segmentation calculation to obtain each scanning window. Corresponding segmentation mask ,use The extracted feature vector of the current frame is (Same method as S2), using the binary mask of the current frame's real image ( The extracted feature vector is (Same method as S2), calculate the layered shape loss. (Method same as S2). After each scan calculation, target re-identification is performed: among all scan window segmentation results, the segmentation mask with the smallest layer shape loss is selected. As a target for re-identification:

[0216]

[0217] And must meet If a matching candidate region is found, the re-identification is successful, the system updates the tracking status to normal tracking, and the settings are updated. and according to Global coordinates are calculated within the scanning window to complete target relocalization. If no re-identified target meeting the criteria is found within the search range, the target is determined to be continuously lost, and the re-identification process continues in the next frame.

[0218] The re-identification mechanism performs a refined search in the target loss area through shape feature matching, effectively solving the problem of re-capture after the target is briefly occluded or disappears, and significantly improving the robustness and sustainability of the tracking system in complex scenarios.

[0219] Example explanation: 1) Occlusion trigger: In frame 0023, the candidate region area is 3 < 5, and Determine occlusion and record the target center coordinates in the previous frame's prediction probability map. (Target center coordinates in frame 0022) triggers scan re-identification; no target meeting the criteria was found in the scan window from frames 0024 to 0035 (the area of ​​all candidate regions within the scan window is <5 or <10). All of these output "area=0, pos=None";

[0220] 2) Re-identification successful: Frame 0036 scanned a window offset by 36 pixels (center (62.837, 24.705)), the segmentation mask area 17 > 5. Furthermore, it exhibits minimal layered shape loss and serves as a target for re-identification. The global coordinates are (62.837, 24.705), and the output is "(scanning) Found at offset 36: area=17, global=(62.83747276688449, 24.705045665775877)".

[0221] 3) Tracking recovery: From frame 0037 onwards, the tracking results are stable (e.g., in frame 0037, the candidate region area is 27 and the target center coordinates are (60.356, 24.372); in frame 0038, the area is 28 and the center is (56.856, 25.943)). All values ​​are <0.8, indicating a return to normal tracking status.

[0222] Based on the specific data and performance of the training and tracking process in the above examples, the infrared target anti-jamming tracking method based on the feature segmentation tracking network (FASTNet) provided in the application exhibits excellent performance in the three core indicators of accuracy, robustness, and real-time performance, fully meeting the needs of infrared target tracking in complex interference scenarios.

[0223] From the training accuracy, the sub-losses of the comprehensive loss function all achieve stable convergence: the Soft-IoU loss gradually decreases from 0.9577 in the initial Epoch 1 to 0.0174 in Epoch 20, meaning that the overlap between the predicted mask and the true mask increases from less than 5% to 98.26%; the energy loss is optimized from 0.8670 in Epoch 1 to 0 after Epoch 10, indicating that the target area prediction error is completely eliminated; the hierarchical shape feature loss converges from 1.0802 in Epoch 1 to 0, proving that the network accurately learns the shape structure features of the target (such as the difference between the predicted shape and the true shape in frame 0076 is only 0.00022). The extreme convergence of the three types of losses fully verifies the network's accurate segmentation capability in all directions of "overlap-area-shape" for infrared targets.

[0224] From the tracking robustness, the network performs particularly well in complex scenarios: in the normal tracking stage (frames 0005-0022), the target area is stable at 17-30 pixels (highly matched with the true target area of 30), and the global center coordinates smoothly transition from (143.5, 37.67) to (111.6, 26.23), fully adhering to the actual motion trajectory of the target; in the occlusion scenario (occlusion triggered at frame 0023 and recovered at frame 0036), even though there are 12 frames without effective targets (output area = 0) in the middle, the network can still accurately locate the target at a shift of 36 pixels (area = 17, global coordinates (62.84, 24.71)) through the scanning and re-identification mechanism; in the multi-interference scenario (frame 0076), the network accurately selects the true target from the three candidate regions (scores 0.8504, 3.2158, and 2.7276), avoiding tracking drift caused by background interference.

[0225] In terms of real-time performance, relying on lightweight network design (input 32x32 local patch, channel number 16-48, no redundant modules), the actual running speed on the experimental hardware (i7-11800H + RTX 3060 Laptop GPU) reaches more than 25fps, meeting the engineering needs of the infrared tracking system for "real-time response" and avoiding the lag or target loss problems caused by the computational complexity of traditional segmentation tracking algorithms.

[0226] In conclusion, the infrared target anti-interference tracking method based on the segmentation tracking network architecture designed in the application has excellent effect, which not only solves the pain points such as occlusion sensitivity and drift of the infrared target in a complex interference scene, but also guarantees the real-time performance, fully verifies the practicability and superiority in the practical application of military reconnaissance, security monitoring and the like, and provides an efficient and reliable technical scheme for the infrared target anti-interference tracking.

[0227] The above merely describes preferred embodiments of the application, and is not intended to limit the application. Various modifications and changes can be made by those skilled in the art based on the spirit and principles of the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. An infrared target anti-jamming tracking method based on a feature segmentation tracking network, characterized in that, Includes the following steps: S1, Construct a neural network model: Input the real images of the previous few frames containing the target region and the real image of the current frame. After performing inter-frame attention enhancement, inter-frame difference enhancement, gradient enhancement and spatial location enhancement on the input image in sequence, the feature enhancement map of the current frame is obtained. The feature enhancement map is then subjected to downsampling, feature perception and feature fusion in sequence, and the predicted probability map of the current frame is output. S2, Define constraints: Generate a comprehensive loss function based on the Soft-IoU loss function, energy loss function and hierarchical shape loss function. The comprehensive loss function is used to constrain the prediction probability map from three dimensions: overlap, energy and shape. S3, Determine the tracking status: Extract several candidate regions from the prediction probability map, calculate the comprehensive score of each candidate region using the Soft-IoU loss function and the hierarchical shape loss function, and determine the tracking status of the neural network model on the target based on the comprehensive score of each candidate region and the feature degradation index. In step S1, the specific steps for obtaining the feature enhancement map of the current frame by sequentially performing inter-frame attention enhancement, inter-frame difference enhancement, gradient enhancement, and spatial location enhancement on the input image are as follows: A blurred mask is generated based on the binary mask of the previous frame's real image, and a temporal correlation is constructed. By subtracting the real images of adjacent frames, several inter-frame difference channels are generated to enhance temporal correlation. Gradient filtering is applied to the real image of the current frame in different directions to construct gradient channels; The planar coordinates of each pixel in the current frame's real image are normalized to obtain the horizontal and vertical coordinate channels. Perform channel concatenation and convolution operations on the real image, the blur mask, the inter-frame difference channel, the gradient channel, the horizontal axis channel, and the vertical axis channel to obtain the feature enhancement map of the current frame; In step S1, the specific steps of performing downsampling, feature perception, and feature fusion on the feature enhancement map in sequence to output the prediction probability map of the current frame are as follows: Residual enhancement and downsampling are performed on the feature enhancement map to generate shallow feature map, mid-level first-class feature map, mid-level second-class feature map, deep first-class feature map, and deep second-class feature map; Multi-scale dilated convolutions are performed on the intermediate-level first-class feature maps and the deep-level first-class feature maps respectively to generate intermediate-level perceptual feature maps and deep-level perceptual feature maps; Cross-gated fusion is performed on the intermediate level two feature map and the upsampled deep perception feature map to generate the first feature fusion map; The first feature fusion map and the middle-layer perception feature map are cross-gated and fused to generate the second feature fusion map. Cross-gated fusion is performed on the upsampled second feature fusion map and the deep second-level feature map after two levels of upsampling to generate the third feature fusion map; Cross-gated fusion is performed on the shallow feature map after residual enhancement and the third feature fusion map after two-level residual enhancement to generate a fourth feature fusion map. Two-stage residual enhancement and convolution operations are performed on the fourth feature fusion map to obtain the prediction probability map; The residual enhancement includes residual convolution and gradient attention enhancement. Gradient attention enhancement includes channel attention enhancement and spatial attention enhancement, expressed as follows: in, Indicates an enhanced object, Indicates the number of channels. Indicates to Perform convolution and output as One channel, This indicates the first batch of normalization. Indicates ReLU function activation. This indicates the second batch of normalization. This represents the output result of the residual convolution. Indicates average pooling. This indicates max pooling. This indicates the first full connection. This indicates a second full connection. This indicates that the Sigmoid function is activated. This indicates the results of channel attention enhancement. Indicates channel splicing. This indicates the results of spatial attention enhancement. This represents the residual adjustment term. Indicates to Perform residual enhancement and output as One channel.

2. The method according to claim 1, characterized in that, In S1, the expression for generating shallow feature maps, intermediate first-level feature maps, intermediate second-level feature maps, deep first-level feature maps, and deep second-level feature maps by performing residual enhancement and downsampling on the feature enhancement map is as follows: in, Represents the feature enhancement map, This indicates that the residual has been enhanced. This represents two-dimensional max pooling. This represents a shallow feature map. Indicates the number of shallow channels. This represents the first-level feature map in the middle layer. This represents the second-level feature map of the middle layer. Indicates the number of middle-layer channels. This represents the deep first-level feature map. This represents a deep secondary feature map. This indicates the number of deep channels.

3. The method according to claim 1 or 2, characterized in that, The expression for cross-gated fusion is: in, This represents the high-resolution feature maps that need to be fused. This represents the low-resolution feature maps that need to be fused. Indicates the number of projection channels. Represents a 1×1 convolution. This indicates the result of high-resolution feature fusion. This indicates the result of low-resolution feature fusion. This indicates that the Sigmoid function is activated. Indicates the gating coefficient. This indicates the features after gating. Indicates channel splicing. This represents a 1×1 convolution with batch normalization and ReLU activation. This represents a feature fusion map.

4. The method according to claim 1 or 2, characterized in that, In S2, the expression for the comprehensive loss function is: in, , , These represent the Soft-IoU loss weights, energy loss weights, and layered shape loss weights, respectively. Indicates Soft-IoU loss. Indicates energy loss. Indicates layered shape loss, This indicates the overall loss.

5. The method according to claim 1 or 2, characterized in that, In S2, the expression for the energy loss function is: in, Represents pixel coordinates, Indicates the current frame's true image in grayscale value at that location The binary mask representing the real image of the current frame is in The value at ∑ represents the summation of pixel values ​​at all locations in the image. Represents normalized real energy. Indicates the current frame prediction probability map in The value at that location, This represents the normalized predicted energy. Indicates the smoothing term. This indicates energy loss.

6. The method according to claim 5, characterized in that, In S2, the expression for the layered shape loss function is: in, The binary mask representing the real image of the current frame is in The value at that point or the predicted probability map of the current frame after binary segmentation The value at that location, This represents the normalized grayscale value of any pixel in the target region of the current frame's real image. Indicates the area of ​​the target region. Indicates the x-coordinate of the center of the target area. Represents the ordinate of the center of the target area. Represents the normalized maximum radius. Represents the center-normalized x-coordinate, Represents the center-normalized ordinate. Represents pixels normalized radius, Indicates the feature vector dimension index. The normalized radius is The first pixel Each feature weight, Indicates use The extracted first eigenvalues ​​of order 1 Indicates use Extracted feature vector of the current frame, Indicates use The calculated first eigenvalues ​​of order 1 Indicates use The calculated first eigenvalues ​​of order 1 Indicates the first Relative difference of each feature Represents the numerically stable term. These represent the loss weights for the first, second, third, and fourth layers, respectively. These represent shape loss at one layer, two layers, three layers, and four layers, respectively. This indicates the loss of layered shape.

7. The method according to claim 6, characterized in that, The specific steps of S3 are as follows: Binarization and connected component operations are performed on the predicted probability map to extract several candidate regions, as shown in the expression: in, Indicates the binarization threshold. This represents the predicted probability map for the current frame. This indicates an 8-neighbor connectivity operation. Indicates the first Candidate regions, Indicates the area of ​​the candidate region. Indicates the area threshold. Represents the set of candidate regions; The comprehensive score for each candidate region is calculated using the Soft-IoU loss function and the hierarchical shape loss function. The expression for the comprehensive score is as follows: in, These represent the weights of the Soft-IoU loss score and the weights of the hierarchical shape loss score, respectively. This represents the target region in the prediction probability map of the previous frame. express Soft-IoU loss, Represents the binary mask of the real image in the current frame. express Layered shape loss, This represents the overall score of the candidate region; When the target is not occluded, the hierarchical shape loss of at least one candidate region is less than or equal to the feature degradation index, and the candidate region with the smallest comprehensive score is selected as the predicted target region. When the target is occluded, the layered shape loss of all candidate regions is greater than the feature degradation index. A neural network model is used to re-track the target until the layered shape loss of at least one candidate region is less than or equal to the degradation index. The candidate region with the smallest layered shape loss value is selected as the predicted target region.

Images